But that may leave VMware, Microsoft, and Citrix users feeling left out. KVM is an efficient hypervisor incorporated into the Linux kernel, but it still represents the smallest installed base of existing virtualization users.
Windows users might feel doubly slighted. Not only is there's no provision for Microsoft's Hyper-V hypervisor in Compute Engine, but Windows workloads themselves are not welcome. That would seem to be an oversight, since Windows and Linux now occupy 80-90% of the enterprise data center, by some estimates.
Even Google's Linux acceptance is limited at this stage. Google infrastructure-as-a-service (IaaS) will take jobs running on CentOS and Ubuntu. Red Hat helped Google architect its cloud to make efficient use of KVM, according to Joe Beda, lead software engineer for Compute Engine, in a talk at the Google I/O developer conference Thursday in San Francisco. Maybe support for Red Hat Enterprise Linux will come later.
Google, with its search engine data center expertise, has the smarts to offer competitive infrastructure. Executives at Amazon Web Services are no doubt warily eying what Google is up to, but AWS's EC2 is not in any immediate danger. Amazon has years of experience in running both Windows and Linux workloads.
[ Want to see how difficult it can be to weigh the value of different cloud service providers? See Why Cloud Pricing Comparisons Are So Hard. ]
One important differentiator did emerge. Google Compute Engine automatically encrypts communications between virtual machines and will let all of a customer's virtual machines talk to each other in a secure fashion, even when they're located in separate data centers.
That's because Compute Engine's three data centers are connected by private communications lines, not the public Internet. Those data centers span three locations: the "central U.S." and two unnamed, but geographically separate, locations on the East Coast, said Beda. Google customers will be able to commission backup copies of virtual machines or disaster recovery systems with some ease because of the secure links in place between them. "If packets come in from an IP address that's your virtual machine, we will guarantee they came from there," said Beda during his talk. Google will charge $0.01 per GB of bandwidth consumed.
Similar arrangements can be made through other service providers, but they may require invoking higher priced private cloud services from the provider and the customer contracting for private line service between locations from network providers.
Google is charging less than Amazon or Microsoft for comparable servers, but it appears to be offering a slightly heftier resource package. For example, its base unit, referred to as n1-standard-1-d, includes 3.75 GB of RAM and 420 GB of local disk for a charge of $0.145 an hour. Amazon's "medium" Linux server is nearest in price at $0.16 per hour, and it includes 3.7 GB of RAM and 410 GB of storage. Microsoft charges $0.24 per hour for a medium-sized Windows server with 3.5 GB of RAM and 490 GB of storage.
What's harder is comparing the CPU power that comes with each instance. Google is introducing the Google Compute Engine Unit (GCEU)--a new unit of measure for virtual machine CPUs--to an already varied and confused field. The virtual CPU of Compute Engine’s base server consists of 2.75 GCEUs.
Google documentation says a GCEU has a defined physical equivalent: If a GCEU is multiplied by 2.75, it's equal to the minimum power of one Intel Sandy Bridge core thread. (Warning: Google documentation thinks the acronym GCEU should have a shorter and cooler version, GQ. It uses both. Amidst already confusing terms, we're given an acronym for an acronym.)
Craig McLuckie, lead product manager, sensing the confusion inherent in the definition, imposed his own description in an interview Thursday: The base server has the rough equivalent of half a Sandy Bridge core; it commands one of the two threads that run simultaneously on the chip, and in some cases it will be able to make use of more than half the chip's resources. As a minimum, it always has the use of one thread, he said. A GCEU is also meant to be a rough equivalent to Amazon's unit of measure of virtual server CPU. That's an EC2 Compute Unit or ECU, which is defined as a single core of a 2007 Intel Xeon running at 1 GHz. Sandy Bridge is a more powerful 2011 chip.
McLuckie said Compute Engine will be accessible to developers and customers through public APIs, the same as Google App Engine is. And he said Google will strive to deliver consistent performance from its service.
An early user, Hamza Kaya, software engineer of Invite Media, said his firm must process 400,000 ad requests a second and was doing so on a well-known IaaS provider. Tests indicated it could run its operation more efficiently on Compute Engine. In two weeks, it was able to move its operations into the Google cloud, where it found each server was processing an average of 650 ad queries per second instead of the previous 350. In each case it commissioned eight-CPU servers, but found it needed 140 in Compute Engine compared to 284 from its previous supplier.
Invite Media systems need to make calls to backend database servers. When those calls take 10 seconds or more, they are interrupted and repeated as a connection error. Its connection error rate dropped from 5% of queries down to 0.5%, said Kaya.
It's too early to say whether the Compute Engine's approach to networking, pricing, and server resource assignment will prove a winning combination in the marketplace. Google has lagged the market in putting together an infrastructure offering and now must find a way to get back in the race. It was only last November that it took its predecessor platform-as-a-service, Google App Engine, out of the experimental preview phase after three years of operation.
But few dispute that Google can bring a decade of experience in search engine operations to the endeavor. It will need to simplify some of the complex messages it appears wedded to in explaining the service. And it will need to find more compelling ways to describe what its infrastructure can do on behalf of an established business. But the growing appetite for plain infrastructure-as-a-service still gives it a chance to do exactly that.
Expertise, automation, and silo busting are all required, say early adopters of private clouds. Also in the new, all-digital Private Clouds: Vision Vs. Reality issue of InformationWeek: How to choose between OpenStack and CloudStack for your private cloud. (Free with registration.)